1d ago

Salesforce releases Procedural Memory Distillation to let models reuse signals from previous training runs for cumulative self-improvement

It distills self-teacher guidance directly into student weights.

——0——
Original post
Rishabh Agarwal#196@AGARWL_OPSalesforce AI ResearchSalesforce AI Research@SFRESEARCH

Can Language Models Remember What They Learn? Introducing Procedural Memory Distillation (PMD): http://sforce.co/4dAjQOu PMD turns model attempts into reusable training memory, conditions a self-teacher on it, and distills the guidance into the student's weights.

2:49 PM · May 28, 2026 View on X
QUOTE POSTSilvio Savarese#796Silvio Savarese@SILVIOCINGUETTA

Post-training has been episode-local for too long. Signal from prior attempts gets discarded after a single update. PMD changes the equation by letting policy and memory co-evolve, turning a model's own training history into cumulative self-improvement. #SystemLevelAI #SelfEvolvingAI

Salesforce AI ResearchSalesforce AI Research@SFResearch

Can Language Models Remember What They Learn? Introducing Procedural Memory Distillation (PMD): http://sforce.co/4dAjQOu PMD turns model attempts into reusable training memory, conditions a self-teacher on it, and distills the guidance into the student's weights.

9:49 PM · May 28, 2026 · 2.1K Views
3:07 AM · May 29, 2026 · 1.1K Views